Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Interpretability for Vision Models via Shapley Value Optimization
Kanglong Fan, Yunqiao Yang, Chen Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with Shapley value auxiliary loss | ViT-B · ImageNet classification | Acc@1: 75.0 -> 76.8 | Improved |
| fine-tune with Shapley value auxiliary loss | ResNet-50 · ImageNet classification | Acc@1: 75.3 -> 76.3 | Improved |
| fine-tune with Shapley value auxiliary loss (ResNet-50) | ResNet-50 · ImageNet classification | FLOPs: 4.1B -> 6.6B | Hurt |
| use Shapley self-explaining training instead of post-hoc TokenTM | ViT-B · ImageNet explanation faithfulness | AOPC: 0.755 -> 0.803 | Improved |
| use Shapley self-explaining training instead of post-hoc FastSHAP | ResNet-50 · ImageNet explanation faithfulness | SaCo: 0.3197 -> 0.5193 | Improved |
| use Shapley training instead of B-cosified | ViT-B · ImageNet classification | Acc@1: 75.3 -> 76.8 | Improved |
| use Shapley training instead of B-cosified | ResNet-50 · ImageNet classification | Acc@1: 76.5 -> 76.3 | Hurt |
| use Shapley training instead of HarsanyiNet | custom CNN · ImageNet-Segmentation | mIoU: 66.03 -> 66.96 | Improved |
| use Shapley training instead of BagNet | custom CNN · ImageNet-Segmentation | mIoU: 66.77 -> 66.38 | No clear effect |
| fine-tune EVA-CLIP with Shapley loss | EVA-CLIP ViT-B/16 · MS COCO zero-shot retrieval (T2I) | R@1: 42.2 -> 42.1 | No clear effect |
| fine-tune EVA-CLIP with Shapley loss | EVA-CLIP ViT-B/16 · Flickr30k zero-shot retrieval (T2I) | R@1: 71.2 -> 70.8 | No clear effect |
| use Shapley training instead of B-cosified | EVA-CLIP ViT-B/16 · MS COCO explanation insertion (T2I) | Insertion R@1: 0.1558 -> 0.1679 | Improved |
Closest papers by meaning
- Learning to Estimate Shapley Values with Vision Transformers (ICLR 2023 · similarity 0.91)
- Gnothi Seauton: Empowering Faithful Self-Interpretability in Black-Box Transformers (ICLR 2025 · similarity 0.88)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.86)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.86)
- Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
